Automated persona vector extraction pipeline — automates away the bespoke curation of → Contrastive Activation Addition (CAA)

Persona-vectors' related-work section places CAA in a short list of techniques it says 'generally required bespoke data curation to obtain contrastive pairs' (persona-vectors, §"7 Related work", p. 12), but that single sentence does not itself spell out what the bespoke curation looked like. Reading it against CAA's own machinery fills that in: CAA's seven target behaviors each needed their own hand-assembled or GPT-4-generated multiple-choice dataset — a custom hallucination dataset distinguishing unprompted from contextually-triggered fabrication, a custom refusal dataset, others drawn from Anthropic's Advanced AI Risk and Sycophancy evaluation sets — with hundreds of contrast pairs curated per behavior before any vector could be built (contrastive-activation-addition, §"3 Method", p. 3). The extraction pipeline replaces all of that per-behavior dataset engineering with one fixed prompt template that needs only a trait name and a short description, run identically for every trait it studies — the three main traits plus four more in Appendix G — with no behavior-specific dataset construction anywhere in the loop (persona-vectors, §"2.1 Generating trait-specific artifacts", p. 3). What CAA's specific curation effort actually entailed is not something persona-vectors elaborates on; the pipeline's real relationship to it — replacing seven separate dataset-building projects with one reusable template — only comes into focus by reaching into CAA's own paper for the detail persona-vectors' one-line characterization skips.